What is an AI company operating system?
A business operating system is a team's shared processes, roles and accountability. An AI company operating system adds a persistent team of AI agents that works from that shared context, skills and memory. Agent Swarm is one you can run yourself: open source, self-hosted, on Claude Code, Codex, or any runtime.
| Approach | What it is | What it organises | Who does the work | What persists |
|---|---|---|---|---|
| Conventional business operating system | A documented way of running a company: processes, roles, meetings and scorecards. Frameworks such as EOS are examples. | People's recurring work and how they are held accountable. | People, in the roles and meetings the framework defines. | The playbook: process documents, role charts and meeting rhythms. |
| AI application framework (LangChain, CrewAI, OpenAI Agents SDK) | A library developers use to build agent applications. | Agent logic inside one application you are building. | Your developers. They wire up state, memory and deployment themselves. | What you build and operate on top of it. |
| Persistent operating swarm (Agent Swarm) | A running team of agents you hand work to. | Work across Slack, Linear and GitHub: a lead agent routes tasks to worker agents in isolated containers. | The agents, with your review gates where you set them. | Shared memory, skills, schedules and review rules, across sessions. |
A swarm can carry out the operating system your team already runs. It does not replace it.
See how Agent Swarm compares to frameworks and platforms on the comparison hub, or read the documentation.
Agent Swarm is the open-source, self-hosted AI company operating system built by Desplega Labs. It is not OpenAI Swarm, OpenAI's experimental educational library, replaced by the OpenAI Agents SDK.
Who it is for
It fits teams that coordinate recurring work across shared context.
- Coding. Pull requests and code review, from issue to reviewed PR, without re-briefing the agent each time.
- Research. Recurring briefs run on a schedule, and each run starts from what the last one learned.
- Operations. Incident triage and scheduled checks follow your documented method.
- Go-to-market. Weekly sales and marketing briefs and content drafts, built on shared context and skills.
See real sessions in the examples.
Limits and boundaries
- You own the infrastructure. Self-hosted with Docker Compose or the Helm chart, so you run, patch and pay for it. The managed Cloud is waitlist-only; see pricing.
- Your model provider has its own terms. You bring your own model keys and harness: Claude Code, Codex, pi, opencode, Cursor, Amp, Devin, or ACP agents. Prompts and context sent to a provider fall under that provider's terms. Agent Swarm does not make provider-side processing private.
- Shared context is not unrestricted access. Agents only reach the integrations and credentials you configure. See security.
It is not the right fit for one-off chats, or for teams with no repeated work worth turning into a skill.